Machine learning approaches for predicting antibiotic resistance genes abundance changes during biological nitrogen removal process.

Wastewater treatment plants (WWTPs) serve as reservoirs for multiple antimicrobial agents (AAs), thereby promoting the risk of antibiotic resistance genes (ARGs) transmission in sewage and sludge during biological nitrogen removal (BNR) processes. And the fate of ARGs is challenging to be deciphered due to the influence of multiple factors and microbial interactions. This study employed four machine learning (ML) models for predicting changes in the abundance of ARGs and mobile genetic elements (MGEs) during BNR processes. Categorical Boosting (CatBoost) model achieved the highest prediction accuracy for predicting ARGs abundance changes (R = 0.843), while the random forest (RF) model outperformed other models in predicting MGEs abundance changes (R = 0.708). Feature importance analysis indicated that MGEs abundance changes was the most important feature for predicting ARGs abundance changes, while two microbial community characteristics (Bacteroidetes and Proteobacteria abundance changes) and two environmental factors (exposure time and pollutant concentration) were identified as critical features for influencing both ARGs and MGEs abundance changes. And the control strategies for reducing ARGs transmission risks were proposed based on Partial Dependence Plots (PDPs) analysis. A user-friendly graphic interface was developed to provide operational guidance for optimizing ARGs mitigation strategies. Overall, this study offered effective ML approaches to evaluate key factors and guide the control of ARGs transmission risks during BNR processes.